Ecological Applications, 19(6), 2009, pp.

15611573 2009 by the Ecological Society of America

Habitat conversion, extinction thresholds, and pollination services in agroecosystems

TIMOTHY H. KEITT1
Integrative Biology, University of Texas, Austin, Texas 78712 USA

Abstract. Parallel declines of wild pollinators and pollinator-dependent plants have raised alarms over the loss of pollination services in agroecosystems. A spatially explicit approach is needed to develop specic recommendations regarding the design of agricultural landscapes to sustain wild pollinator communities and the services they provide. I modeled pollination services in agroecosystems using a pair of models: a stochastic individual-based simulation model of wild pollinators, pollinator-dependent plants, and crop pollination; and a set of coupled difference equations designed to capture the nonspatial component of the simulation model. Five spatially explicit models of habitat conversion to crops were simulated, and results for pollination services were compared. Mean-eld behavior of the simulation model was in good agreement with analysis of the difference equations. A major feature of the models was the presence of a cusp leading to loss of stability and extinction of pollinators and pollinatordependent plants beyond a critical amount of habitat loss. The addition of pollen obtained from crop visitation caused a breakdown of the cusp preventing extinction of pollinators, but not of wild pollinator-dependent plants. Spatially restricted foraging and dispersal also altered model outcomes relative to mean-eld predictions, in some cases causing extinction under parameter settings that would otherwise lead to persistence. Different patterns of habitat conversion to crops resulted in different levels of pollination services. Most interesting was the nding that optimal pollination services occurred when the size of remnant habitat patches was equal to half the mean foraging and dispersal distance of pollinators and the spacing between remnant patches was equal to the mean foraging and dispersal distance. Conservation of wild pollinators and pollinator-dependent plants in agroecosystems requires careful attention to thresholds in habitat conversion and spatial pattern and scale of remnant habitats. Maximization of pollination services was generally incompatible with conservation of wild pollinator-dependent plants. My prediction is that pollination services will be maximized by providing islands of nesting habitat where interisland distance matches mean foraging distances of wild pollinators.
Key words: agroecosystems; Allee effect; conservation; difference equations; ecosystem services; habitat conversion; habitat fragmentation; landscape model; mutualism; pollination services.

INTRODUCTION Human alteration of landscapes for food production affects ecosystems worldwide and is possibly the largest single component of global change (Matson et al. 1997, Vitousek et al. 1997, Tilman et al. 2001). The resultant agroecosystems are often highly fragmented with areas of intensive cultivation interspersed among remnant patches of original habitat. The close juxtaposition of seminatural remnants to intensively managed lands creates a strong opportunity for spill over effects where ecological interactions extend across habitat boundaries to affect ecosystem function in adjacent patches (Rand et al. 2006). In a recent synthesis, Kremen et al. (2007) evoked the phrase Mobile Agent-Based Ecosystem Services (MABES) to describe situations in which animal vectors
Manuscript received 22 January 2008; revised 17 October 2008; accepted 5 December 2008. Corresponding Editor: E. A. Newell. 1 E-mail: tkeitt@mail.utexas.edu

propagate ecological functions across habitat boundaries. Here services refers to the subset of these interactions that directly or indirectly affect human welfare. A multitude of services can be imagined ranging from animal seed dispersal to predatory control of pest species (Kremen and Ostfeld 2005). Among these, pollination services hold particular importance. More than one-third of global agricultural output depends on animal pollinators (Klein et al. 2007). The value of wild-pollinated agricultural products may range as high as U.S.$3 billion per year within the United States (Losey and Vaughan 2006), and as high as $127 billion per year worldwide (Costanza et al. 1997). There is growing concern over the large-scale conversion of landscapes to agricultural and other human land uses and its impact on pollinator communities (Kearns et al. 1998, Klein et al. 2003, Tscharntke et al. 2005, Biesmeijer et al. 2006, Cane et al. 2006, Pauw 2007, Steffan-Dewenter and Westphal 2008). Pollen supplementation experiments indicate many natural plant

1561

1562

TIMOTHY H. KEITT

Ecological Applications Vol. 19, No. 6

communities experience pollen limitation (Cunningham 2000, Donaldson et al. 2002, Ashman et al. 2004, Aguilar et al. 2006, Wagenius et al. 2007). Pollen limitation is likely to be even more severe in agroecosystems because crops are often grown in large blocks inhospitable to nesting of wild pollinators. In such cases, the large travel distances required to reach crops in the interior of elds may discourage visitation (Cresswell et al. 2000, Priess et al. 2007). Pollen limitation has the potential to affect not only crop production, but also wild pollinator-dependent plants in remnant habitat fragments (Aizen and Feinsinger 1994, Matsumura and Washitani 2000). The conservation implications of pollen limitation have led to calls to determine exact thresholds of habitat conversion leading to plant and pollinator extinctions (Ashman et al. 2004). Domesticated bee pollinators have long been used to enhance pollination services to crops, a practice that can offset losses of wild pollinators. However, reliance on a pollinator monoculture carries strong risks. Populations of managed and feral honey bees for example have recently experienced strong declines across North America (Allen-Wardell et al. 1998) increasing costs to farmers (Kevan and Phillips 2001). The low genetic diversity of domestic pollinators may expose them to high risk of epidemic disease and there is evidence to suggest that a long history of episodic declines in domestic pollinators may be due to pathogen outbreaks (Wilson and Manapace 1979a, b, Cox-Foster et al. 2007, Pettis et al. 2007, Stokstad 2007, Winfree et al. 2007). The developing view of unfavorable economics of domesticated pollinators, along with biodiversity concerns, has focused attention of the problem of restoring wild pollinators to agroecosystems (Kremen et al. 2002, Pywell et al. 2005, Morandin and Winston 2006, Winfree et al. 2007). Enhancement of wild pollination services must have a strong spatial component because of the combined actions of metapopulation dynamics and spillover effects. Hence there is a need to consider in detail the consequences of different patterns and scales of habitat conversion in a spatially explicit individualbased context in order to better understand the impacts of land management decisions on wild pollinators and pollinator-dependent plants (Brosi et al. 2008). I consider models of MABES in agroecosystems and specically focus on how patterns of habitat conversion relate to the persistence of wild pollinators and pollinator-dependent plants in a fragmented landscape. I also model the spillover of pollinators into cultivated areas and evaluate the pollination services provided under different scenarios of habitat conversion. Reasonable models of pollination in agroecosystems can be arbitrarily complex. Additional complexity, however, carries considerable costs in terms of data requirements and difculty of interpretation and generalization. I intentionally chose a highly simplied modeling framework as being most productive in setting a formal basis for understanding landscape aspects of pollination

services. The framework permits both stochastic simulations and equivalent deterministic mathematical representations providing a form of cross validation, yet can be extended to explore more complex scenarios as needed. MATERIALS
AND

METHODS

My results are based on a pair of models: one a detailed individual-based simulation and the other composed of a set of coupled difference equations. The models update the total density of plant populations (P), pollinator populations (N ), and of seeds in a seed bank (S ). As the emphasis of the modeling was on spatial aspects of habitat conversion, details of reproductive life histories were omitted (the only constraint being contact between a plant and a pollinator). Hence one can equivalently assume P and N represent single generalist populations or guilds of ecologically equivalent species. (Generalizations of the model to complex communities are considered in Discussion.) The difference equations were constructed to capture the mean-eld behavior of the simulation model, as occurs when starting and ending points of individual movements are spatially independent. The framework follows closely the yuccamoth model of Wilson et al. (2003), but departs in signicant aspects as described below. I rst describe the details of the individual-based model and then derive the mean-eld equations. Simulation model Plantpollinator dynamics were simulated on an articial landscape composed of a grid of cells. A fraction K of these cells could be occupied by individual plants; the remaining portion of the landscape was assumed empty or occupied by crop plants. As in Wilson et al. (2003), I recorded all population measurements in terms of individuals per landscape grid cell. Plants were restricted to no more than one individual per cell, whereas density of seeds and pollinators could grow as long as recruitment exceeded mortality (see Eqs. 24). Pollinators were assumed to be central-place foragers and visit a xed number of sites per season. Pollinator density was equated to the density of nest sites, which could without loss of generality represent a solitary foraging female or a nest with one or more reproductive females and accessory foragers. It was further assumed that the density of male pollinators was never limiting in pollinator reproduction and that male pollinators contributed nothing to pollination. Newly emerged reproductive females could establish new sites after dispersal away from the natal site. The scheme most closely approximates the ecology of bee pollinators although it may capture certain aspects of other pollinator systems. Simulations were run in one of two modes, either mean eld or spatial. When run in mean eld mode, individuals moved randomly without regard to distance or direction such that a foraging or dispersing

September 2009 TABLE 1. Parameter descriptions.

POLLINATION SERVICES IN AGROECOSYSTEMS

1563

Parameter h / s c dS a dP k m dN

Description

Default value

seed production probability 1 ower visits per nest per season 100 portion of season owers are available 0.5 germination fraction 0.5 seed mortality 0.1 seedling survivorship 1 plant death rate 0.2 nest production probability from plant visit 0.2 nest production probability from crop visit 0.2 nest death rate 0.5

step 4). Those landing in the fraction X of the landscape suitable for growth establish a new nest site. 6) Populations censused. Densities of seeds, plants, and pollinator nests are recorded along with information on crop pollination. Notice that termination of the time step occurs at the beginning of the simulated growing season, just after recruitment, rather than at the end of the season. This choice simplied the derivation of the mean-eld equations. Mean-eld equations I constructed a deterministic mean-eld model as a benchmark for comparison to the individual-based simulations. The model updates stocks (of seeds S, plants P, and pollinator nests N ) over a seasonal time step Dt by computing frequencies of different events related to recruitment and mortality. The frequency of independent events can be computed as the probability of an event multiplied by the number of trials. If the locations of events are chosen at random, the probability of m events occurring within a cell is given by the Poisson distribution p(m; x) (xmex)/m! where x is the number of events per cell. The probability of one or more events is then p(m . 0; x) 1 ex. With these preliminaries, I dene the deterministic, mean-eld system: DS hP 1 e/sN c1 dS DtS dS S Dt h i DP a 1 ec1dS DtS K 1 dP DtP dP P Dt DN XkP m1 K 1 ebsN dN N Dt 2

individual could visit any cell in the landscape with equal probability independent of the individuals current location. When run in spatial mode, movement direction was drawn uniformly on the circle, and movement distance X was drawn from the negative exponential: Pq X x qeqx 1

where 1/q is the mean dispersal distance. To simplify the modeling, I always used the same function for both foraging and dispersal (i.e., dispersal q foraging q). The negative exponential was chosen based on results of eld studies showing exponential decline of pollinator visits with increasing distances (Ricketts et al. 2008). Periodic boundary conditions were imposed such that movement beyond the edge of the grid wrapped around to the opposing side. To simplify linkage of the simulation model to the difference-equations, a seasonal cycle was dened with events (pollination, recruitment, mortality, and so forth) occurring in discrete phases. The following events occurred during each simulation time step: 1) Flowering and pollination. Each plant owers with probability h after which pollination occurs. Each nest generates bs cell visits within the owering period s. Pollination occurs on the rst pollinator visit. 2) Seed and larvae production. Each pollinated plant produces a single seed. Each pollinator making a rst visit to a ower causes the production of a new larval queen with probability k (if a wild plant was visited) and m (if a crop plant was visited). Subsequent pollinator visits have no effect on either plants or pollinators. 3) Seed, plant, nest mortality. Background mortality is assessed on each of the stocks at rates dS, dP, and dN, respectively. 4) Seed dispersal and germination. Seeds are displaced in random directions with distances governed by Eq. 1. Those landing in the fraction K of the landscape suitable for growth germinate at rate c. For any habitat cell not currently occupied by an adult and receiving one or more germination events, a single seedling recruits to the adult population with probability a. 5) Nest recruitment. Larval queens disperse according to Eq. 1 (using the same value of q as was used in

where K is the proportion of the landscape available to plants and X is the fraction of the landscape suitable for pollinator nesting. Denitions and default values of other parameters are provided in Table 1. Terms in Eqs. 24 compute frequencies of events per unit area (again scaled to the area occupied by a single plant). For example, the rst term in Eq. 2 computes the number of seeds recruited per unit area over the period Dt. Frequency of owers per unit area is given by hP. The probability of one or more visits to these owers is given by 1 e/sN where per area visitation rate is computed as the number of pollinator visits per nest per time (/) multiplied by the time period over which pollination occurs (s) multiplied by the number of nests per unit area (N ). Frequencies of other events were computed analogously. The deterministic model is still relatively complex with 12 parameters. Model complexity can be substantially reduced through judicious rescaling of time and state variables. Making the substitutions t 0 tdP, S 0 c(1 dPDt)S, P 0 (1 dPDt)P, and N 0 /sN yields (after dropping primes) the eight-parameter model:

1564

TIMOTHY H. KEITT TABLE 2. Scaled model parameters. Parameter u hc1 dS Dt dP 1 dP Dt c1 dS Dt dS dP a1 dP Dt dP k/s dP 1 dP Dt m/s dP dN dP Description seed production

Ecological Applications Vol. 19, No. 6

Default value 2.8125

seed germination and mortality

2.75

seedling survival

nest production

62.5

nest production from crop visitation

50

nest mortality

2.5

Note: Variables are dened in Table 1.

DS uP1 eN qS Dt DP n 1 eS K P P Dt DN XlP e1 K 1 eN xN Dt

exponential expressions that yield little additional insight and is thus omitted from further discussion. As the model cannot be solved in closed form, equilibrium solutions were obtained numerically using standard methods (Brent 1973). Numerical iteration of the mean-eld equations closely matched simulation results obtained from the discrete stochastic model (Fig. A1). Simulated landscapes Five separate models were used to simulate habitat conversion (Fig. 1). In the island model, patches of original habitat are evenly distributed across a matrix of converted habitat. The moat model does just the opposite creating islands of converted habitat surrounded by a moat-like strip of unconverted habitat. Two wavelet-based models were constructed based on the

where the aggregated parameters are dened in Table 2. In the rescaled model, S is equivalent to the density of germination events, P is the post-mortality plant density, and N is the density of pollinator visits. It may be noted that the equilibrium solution to Eqs. 57 can be collapsed into a single expression only involving nest density. While the collapsed form is convenient for numerical analysis, it involves a large number of hyper-

FIG. 1. Landscape models. Lighter cells represent unconverted habitat; black areas have undergone habitat conversion. In each model, 75% of the area is converted. Grid sizes are 256 3 256 cells: (A) 4 3 4 island landscape, (B) wavelet-synthesized fractionalBrownian landscape with scaling exponent b 1.5, (C) wavelet-synthesized Gaussian landscape with correlation length r 32 grid cells, and (D) spanning-tree landscape with eight-cell nugget.

September 2009

POLLINATION SERVICES IN AGROECOSYSTEMS

1565

method described in Keitt (2000). The rst is the bwavelet model where variance scales as Varw; f } f b and w are wavelet coefcients and f is frequency or the inverse scale of analysis. The b-wavelet model ranges between random white noise (b 0) to strongly clustered (b 3). Intermediate values (1 b 2) yield so-called fractal landscapes where cluster sizes vary from a minimum approaching a single grid cell to a maximum approaching the size of the entire grid. The parameter b can thus be thought of as a condensation measure that determines the extent of fragmentation for a particular value of K. The second wavelet model is the Gaussianwavelet model: Varw; r}e2r
2 2

where r determines the characteristic patch size. Unlike the b-wavelet model, the distribution of patch sizes in the Gaussian-wavelet landscapes are strongly modal rather than hierarchical. The spanning tree model derives from graph theory and is described in more detail in Keitt (2003). In this model, the nugget parameter denes the minimum feature width. Cells in a nugget must be either converted or unconverted and cannot be further subdivided. The spanning-tree method connects all nuggets in the grid by (1) growing a minimum spanning tree while breaking nearest-neighbor ties randomly, and then (2) pruning leaves randomly until the desired extent of habitat conversion is reached. The resulting habitat remnant spans much of the grid yet remains fully connected. This is a useful contrast to the other stochastic landscape models where habitat loss and fragmentation covary. Model simulations Simulation runs involved choosing parameters, setting initial conditions and iterating the model until no change was detectable in average numbers of seeds, plants and nests between successive 50-step intervals. In all simulations, a grid size of 256 3 256 was used to ensure results were not unduly biased by nite-size effects. Initial placements of seeds, plant, and nests were uniform random. Pollination services were measured as the density of crop plants receiving one or more visits during a simulated season (one step) averaged over the nal 50 steps of a simulation run. Note that density was measured relative to the entire simulated landscape. Hence if 10% of the landscape was in crops and all crop plants received pollinator visits, then pollination services over the entire landscape was recorded as 10%. Pollination services measured in this way are directly related to potential crop yield accumulated over the entire landscape. In order to determine properties of landscapes that resulted in the greatest pollination services (see Fig. 6 and Appendix C: Fig. C1), I ran a series of simulations for each of the landscape models described above. For a

given landscape model, I then varied the amount of habitat conversion (x-axis in Fig. 6 and Appendix C: Fig. C1) and the degree of fragmentation or clustering (y-axis in Fig. 6 and Appendix C: Fig. C1). At each level of habitat conversion and fragmentation, I ran 10 replicate simulations each with a different random landscape drawn from the appropriate model (described in Materials and methods: Simulated landscapes) and with random initial placement of seeds, plants, and pollinators. Maximum pollination services was then determined by recording those parameter values (amount of habitat conversion and fragmentation) that resulted in the greatest pollination services averaged over the 10 replicate simulation runs. In Fig. 6 and Appendix C: Fig. C1, the parameter combination that resulted in the greatest pollination services for a given landscape model is indicated by a square symbol. In addition, I added contour lines to aid in visualizing pollination services over the full range of landscape parameters simulated. To do this, I tted a thin-plate spline surface to the results using the Tps function from the elds package in R (R Development Core Team 2008) and subsequently generated height contours for plotting. RESULTS Persistence of pollinators and wild plants under habitat loss Qualitative dynamics and stability were further probed via isocline analysis of the deterministic system (Fig. 2). Plant density dependence combined with increasing pollination failure at very low plant densities generated bistable dynamics with upper and lower stable points separated by an unstable (saddle-point) node. Trajectories simulated in the phase plane (dotted lines) conrmed this observation. For the parameters considered here, the unstable extinction threshold occurred at very low plant and pollinator densities. With habitat loss, the extinction threshold was much higher. At 50% habitat conversion (Fig. 2B), both the upper stable solution and the unstable point occur at relatively high seed, plant and pollinator densities and are relatively close together. In this situation, even small disturbances could potentially push the system below the unstable point leading to extinction. The effect of habitat conversion was far less severe when pollinators were allowed to nest in the converted habitat (X 1, gray line in Fig. 2B). The bistable pattern was however maintained. Simulated trajectories show attraction to stable points and repulsion from the unstable node. With 40% habitat conversion (Fig. 2C), the isoclines completely separate indicating a cusp catastrophe. Once the stable and unstable nodes meet, the single remaining solution is zero density for both populations. Nesting in converted habitat averted the catastrophe, but it is clear that further habitat reduction would eventually destabilize this scenario as well.

1566

TIMOTHY H. KEITT

Ecological Applications Vol. 19, No. 6

FIG. 3. Bifurcation diagrams showing cusps at critical value of habitat conversion. The gray lines show results where pollinators were allowed to nest in the converted landscape. The diagrams are for default parameter values with no crop pollen (e 0).

FIG. 2. Pollinator and plant isoclines projected onto the surface of the seed bank isocline. Overlays show simulation results. Solid circles show initial conditions. The dotted lines show subsequent trajectories. Open circles correspond to pollinator nesting constrained to unconverted habitat. Gray isoclines and circles give results where pollinators can nest anywhere. Lambda (K) is the fraction of the landscape that remains as wild, unconverted habitat; DP/Dt and DN/Dt are seasonal growth increments of plant and pollinator populations; X is the proportion of the landscape suitable for

A more global view of model dynamics is provided by constructing a bifurcation plot showing the pattern of stable and unstable solution as model parameters are varied. Bifurcation plots of the plant-pollinator model reiterate the cusp dynamics at a critical level (moving left to right) of habitat conversion (Fig. 3). The cusp occurs where the unstable branch (dashed lines) meets the stable branch (solid lines) eliminating any non-zero

pollinator nesting. (A) No habitat conversion; (B) 50% habitat conversion; (C) 60% habitat conversion. Circles are shown only every ve simulation steps in panels (B) and (C) so as not to obscure the trajectory. The diagrams are for default parameter values with no crop pollen (nest production from crop visitation e 0).

September 2009

POLLINATION SERVICES IN AGROECOSYSTEMS

1567

solution. Sensitivity of the threshold to changes in individual parameters is shown in Appendix B: Fig. B1. The general conclusions of the sensitivity analysis are that the plantpollinator system was more robust to habitat loss when reproductive rates were higher and less robust to habitat loss when mortality related parameters were increased. Position of the cusp was more sensitive to plant and seed related parameters (/, n, q) and less so to parameters governing nest density (l, x). Fig. 4 compares spatial and mean-eld simulations. In this scenario (50% habitat loss), plant and pollinators survived in the mean-eld case, but went extinct in the spatial simulation. Notice however the timescale over which extinction occurred: nearly 600 seasons pass before the collapse. The extended transient behavior of the system is a combined result of positive feedback and spatially localized interactions. As shown in Fig. 4B, large voids opened in the spatial simulations. Because of positive feedback between plants and pollinators, these voids grew slowly until consuming the entire landscape. The slow progression to collapse has important implications for conservation as indicators of instability may only appear after a long period has passed after habitat conversion. The inuence of crop pollen on pollinator and wild-plant persistence When pollinators forage for crop pollen in addition to wild-plant sources, the model dynamics were considerably different (Fig. 5). Under this scenario, as habitat is converted to croplands, the pollinator receives a guaranteed pollen source that is not itself dependent on the pollinator for reproduction. As a result, pollinator populations may be sustained at low wild-plant density or even after wild-plant populations have gone extinct. At moderate extents of habitat conversion, the increased density of pollinators (owing to crop foraging) can cause a breakdown of the bistable pattern, leaving behind a pair of saddle-point nodes (Fig. 5A and shaded regions in Fig. 5B, C). In the bistable condition, zero density is locally stable for both species. Interestingly, with intermediate habitat conversion to crops, the zerodensity solution for wild plants becomes conditionally stable. Once the pollinator has successfully colonized (surviving on crop pollen), the wild plants can subsequently also invade because they have a preestablished source of pollination. Simulation results bear this out (dashed trajectory Fig. 5A). With further habitat conversion, wild plants go extinct (Fig. 5B). Pollinator nesting in the converted habitat has only a small effect on wild-plant persistence (gray lines in Fig. 5B). Loss of wild pollen sources and nesting habitat has a negative impact on pollinators reducing density by roughly half the preconversion density (X K, Fig. 5C). Pollinators were able to persist on crop pollen until ultimately, loss of nesting habitat drives them extinct. When nesting was allowed in converted habitat (gray lines), pollinators showed little

FIG. 4. Comparison of nonspatial (random visitation) and spatial dynamics (q 0.1). Habitat conversion was set at 50%. Location of converted cells was chosen at random. Pollinator nesting was not allowed in the converted habitat. (A) Time series of spatial (black lines) and nonspatial (gray lines) dynamics. The upper, middle, and lower trajectories are seed, plant, and pollinator densities, respectively. (B) Snapshot of spatial conguration with localized movement at around 400 time steps; units are cell location out of 256 3 256 cells. Parameters were set at their default values as specied in Table 2.

effect of habitat conversion until the point of wild-plant extinction, and then increased in density as more crop area was added. The inuence of habitat geometry on pollination services Alternative patterns of habitat conversion to crops had signicant effects on pollination services. In the case of the Gaussian-wavelet landscape model (Fig. 6), increasing average patch size generally had a positive effect on plant and nest densities. Uncorrelated landscapes (i.e., small r) produced results close to the meaneld predictions whereas longer range autocorrelation (larger patch sizes) boosted plant and nest densities

1568

TIMOTHY H. KEITT

Ecological Applications Vol. 19, No. 6

FIG. 5. Isoclines and bifurcation plots with crop pollination. (A) Pollinator and plant isoclines projected onto the surface of the seed bank isocline. Two trajectories are shown: open circles mark points along the trajectory after introduction of pollinators, but prior to reintroduction of wild plants; gray circles mark points along the trajectory after reintroduction of wild plants. (B) Plant bifurcation plot. Gray lines give results when pollinators nest in the crop. The light-gray region indicates where the plant can invade subsequent to the pollinator. Bistable dynamics occur to the left of the light-gray shaded region. (C) Pollinator bifurcation plot. In the light-gray shaded region, stability of solutions depends on whether the plant has colonized or not. All parameters are at default values.

because of reduced fragmentation. Small values of r resulted in near-mean-eld behavior because the resultant fragment sizes were considerably smaller than the 16 cell-width dispersal and foraging distances leading to a more-or-less random pattern of cultivated and uncultivated fragments within the foraging and dispersal range of the pollinators. Note that it is unnecessary to vary both dispersal distance and patch size as it is their ratio that matters, so one can be held constant while the other varies. For completeness, additional simulations were run with much longer and much shorter dispersal and foraging distances (results not shown). For a xed landscape pattern, longer-range dispersal tended to push the results more toward the mean-eld limit whereas shorter-rage dispersal tended to exacerbate fragmentation effects as well as limit the number of owers accessible to pollinators. Pollination services varied considerably with changes in amount of habitat conversion and landscape autocorrelation (Fig. 6C). Several interesting results were found. First, creating large, uninterrupted crop elds diminished pollination services by a signicant amount despite having a positive effect on the density of pollinators. This result was repeated across all simulations and all landscape models. Although pollinators did better in these landscapes, spillover into crops occurred less frequently resulting in less crop pollination overall. Second, as one would expect, total density of pollination services increased with increasing conversion of the landscape to crops. In the Gaussian-wavelet model, maximum pollination services occurred at approximately 90% habitat conversion, resulting in 83% of the landscape producing harvestable crop. Another interesting result was the apparent matching of the scale of habitat fragmentation and mean foraging and dispersal distance. The greatest pollination services occurred when the autocorrelation parameter (r) was set to 16 cell widths, the same value as the mean foraging and dispersal distances used in the simulations. Further analysis of the Gaussian-wavelet landscapes with r 16 revealed that this value produced patches roughly eight cells across and separated by approximately 16 cell widths. Some variation in patch sizes and separation did occur owing to random variation. This result nonetheless suggests a general pattern of matching foraging and dispersal distances to mean inter-patch distances. Simulations using four additional landscapes (see Fig. 1) generally agreed with results from the island landscape (see Appendix C). DISCUSSION Landscape pattern and scale of fragmentation clearly have a large potential to inuence pollination services in agroecosystems (Fig. 6 and Appendix C: Fig. C1). The principal mechanism at work is a tension between maximizing spill-over of pollinators into crop areas and providing sufciently large and interconnected blocks of nesting habitat. The models presented here

September 2009

POLLINATION SERVICES IN AGROECOSYSTEMS

1569

establish a theoretical basis for understanding the dynamics of pollinators, pollinator-dependent plants, and pollination services in fragmented landscapes. A signicant feature of the models is the presence of a cusp at a critical level of habitat conversion beyond which pollinator and pollinator-dependent plant populations are expected to collapse (see also Amarasekare 2004). While this result may seem obscure from the management perspective, understanding the potential consequences of the nonlinear dynamics governing the collapse is of critical importance (Scheffer and Carpenter 2003). The key lesson is that as habitat is converted, the domain over which the plantpollinator system remains intact is reduced. As a result, even small disturbances may shift the balance toward collapse of wild plants and pollinators in remaining fragments. Providing nesting habitat in converted areas delays, but does not eliminate this effect unless the converted habitat also provides a sustained pollen source. The implication of a reduced domain of stability is that a shift from a sustainable pollinator guild to an extinction trajectory could occur with little warning. The reason is that system dynamics tend to be very slow near isoclines (Hastings 2004). If a plantpollinator mutualism is pushed only slightly past the unstable equilibrium such that it is now attracted to the zero-density solution, it may take a very long time to actually reach extinction. This is because the rate of repulsion away from the unstable solution varies exponentially with distance in the phase space (Seydel 1988). Near the unstable solution deaths will only slightly outnumber births and the system may respond so slowly that the decline in density goes unnoticed. Simulation results bear this out (e.g., Fig. 2C). The effect is exaggerated with localized spatial interactions (Fig. 4). Local regions of high plant and pollinator densities will tend to persist for many generations even when the system is globally moving toward extinction. Under monitoring scenarios, two observations would be indicative of a slide toward eventual extinction: an acceleration in the rate of decline as the plant and pollinator populations uncouple and gradually expanding voids absent of both plant and pollinator. Flowering crops may present a signicant resource to attendant pollinator guilds (Westphal et al. 2003). Pollen supply from crops while not immediately dependent on pollinators, can inuence pollinator abundances and pollinator interactions with wild plants. An interesting scenario occurs in the region where the extent of habitat conversion to crops is sufcient to allow pollinators to persist solely on crop pollen, but is not yet extensive enough to cause extinction of wild plants (shaded region in Fig. 5). In this case, crops may act as apparent mutualists (or perhaps apparent commensalists) with wild plants by increasing pollinator density and indirectly wild plant recruitment. The net effect is a breakdown of the Allee effect: as long as pollinators are present, wild plants can colonize regardless of their

FIG. 6. Simulation results using Gaussian landscapes. Autocorrelation (r) ranged from 2 to 128 increasing by powers of 2. Mean dispersal and foraging distances were 16 cell widths (q 0.0625). Simulation grids were 256 3 256 cells. Results were averaged over 10 simulations. Standard errors were negligible and are omitted. (A) Plant density as a function of habitat conversion (1 K). The bold lines show results for the smallest and largest values of r (circles, r 2; diamonds, r 128). The thick gray line gives the mean-eld prediction. (B) Pollinator nest density as a function of habitat conversion. Symbols are as in panel (A). (C) Contour plot showing pollination services as a function of autocorrelation and habitat conversion. The average dispersal and foraging distance is indicated by a gray line. The location of maximum average services (83%) is indicated by a square symbol. All parameters are at default values.

1570

TIMOTHY H. KEITT

Ecological Applications Vol. 19, No. 6

initial density. Interestingly, this may be an effective strategy to recover wild plants and pollinators in areas where habitat conversion has caused widespread declines. Absent bistable dynamics, pollinators should easily spread into mixed cultivated landscapes (provided sufcient nesting habitat in addition to pollen resources) and subsequently sustain reinvasion of pollinatordependent wild plants. This study suggests a number of general management and conservation recommendations. First, mutualistic species are disproportionately sensitive to habitat fragmentation owing to positive feedback and resulting survival thresholds. The model suggests that extinction thresholds for plantpollinator systems occur at 5060% habitat loss depending on whether pollinators nest in converted habitat. If crops planted in the converted habitat provide pollen, the threshold for pollinators can be extended above 90%, or may disappear entirely if pollinators nest in cultivated areas. At this level of habitat conversion, I predict that pollinator-dependent plants will go extinct in remnants owing to loss of metapopulation collapse. However, precise estimates of the extinction thresholds cannot be generalized to all situations as the outcome will depend critically on biological details of particular systems. For example, Ricketts (2004) found persistence of pollinator services in a coffee plantation despite a high percentage of habitat conversion (approximately 85%) and a lack of nesting in the cultivated area. It is also important to consider habitat alteration that may occur beyond the boundaries of a particular management unit. Simulations in this study were run with periodic boundaries to emulate a uniform percentage of habitat conversion across the landscape. In practice, patterns of habitat conversion may change abruptly from one parcel to the next and these boundary effects should be taken into account in any real planning scenario. Nonetheless, as a general rule conservation efforts should pay close attention to trends in extent and pattern of habitat conversion as early as possible and certainly as soon as crop land becomes the dominant land cover. Additionally, in cases where extensive habitat conversion has already occurred, restoration of wild habitats should be considered, however habitat restoration alone may not be sufcient to restore pollination services owing to the Allee effect discussed above. After restoration, an active reintroduction program may be needed to attain a sustainable pollinator community. One of the more intriguing and possibly the most general result of this study is the importance of matching the scale of habitat pattern to scales of pollinator foraging and dispersal. In both the Gaussian-autocorrelation and island landscape models, I found that a habitat-gap distance equal to the mean dispersal and foraging distances was optimal for pollination services in the gap areas. To my knowledge the rst spatially explicit theoretical result dictating a specic recipe to optimize pollination services in agroecosystems (see also Brosi et

al. 2008). Islands of nesting habitat should be spaced apart a distance equal to the average foraging and dispersal distances of pollinators. Island widths should be one-half of the distance between patches. Further studies need to determine whether alternative dispersal functions (e.g., gamma, Gaussian, heavy-tailed, and so on) will alter the outcome. Also, the particular recipe might change when combined with other strategies, such as attracting pollinators to eld margins (Pywell et al. 2005, Carvell et al. 2007, Olson and Wackers 2007). Fortunately, it should be relatively easy to test the scale-matching hypothesis experimentally by allocating nesting habitat in different patterns and assessing pollination limitation in the eld. Recently, Brosi et al. (2008) considered optimal size and spacing of pollinator reservoirs within an agricultural landscape. Brosi et al. simulate pollinator dynamics along one-dimensional transects and use optimization methods to discover congurations that maximize potential crop yield. Although the goals of our work are somewhat different, there are many interesting areas of both agreement and disagreement in results. A general area of agreement is the importance of dispersal and foraging scales when designing landscapes. Clearly very long-range dispersal negates any congurational effects as pollinators can reach any point in the landscape: all that is required is sufcient habitat to maintain the pollinator population somewhere in the landscape. Shorter dispersal distances generally require distances between crops and reservoirs to be closer. One of the interesting ndings of Brosi et al. is solutions involving both large and small clusters of habitat for pollinators: large blocks of habitat to buffer metapopulation effects and smaller dispersed blocks to spread pollinators around. Generally, hierarchical patch-size models did poorly under scenarios considered in this paper. The discrepancy may have to do with particular assumptions driving the models, i.e., one dimension vs. two dimensions and boundary conditions. I consider a very general landscape with periodic boundaries whereas Brosi et al. focus on a single management unit where foraging beyond the management boundary is discounted. Brosi et al. also separate foraging from dispersal distances and it appears that short dispersal distance combined with larger foraging distances tends to favor larger clusters of pollinator habitat. While the results of Brosi et al. and those presented here form a basis for theoretical understanding of pollinators in agricultural landscapes, major challenges remain. Two areas warranting further consideration are the inuences of pollinator foraging behavior and pollinator community composition on pollination services. The models presented here are dramatically simplied versions of real agricultural landscapes. While these simplications resulted in a more tractable framework for understanding effects of habitat conversion on pollination services, there are a number of areas in which the models could be protably extended. First and foremost would be to incorporate life history variation among plants and pollinators to more accurately reect

September 2009

POLLINATION SERVICES IN AGROECOSYSTEMS

1571

the complexity of natural communities. Species embedded in plantpollinator interaction networks express a range of adaptive strategies from generalist to highly specialized (Pellmyr and Thompson 1996, Ringel et al. 1996, Waser et al. 1996, Johnson and Steiner 2000, Bascompte et al. 2003, Vazquez and Aizen 2004), and it would be interesting to incorporate complex interaction structure into spatially explicit modeling of pollination services (see Fortuna and Bascompte [2006] for a spatially explicit model of fragmentation effects on pollinator networks). Another effect of life history variation would be to increase the variance in dispersal and foraging distances. A number of studies have demonstrated strong species-specic variation in foraging distances (Klein et al. 2008, Kohler et al. 2008). Variation in foraging distances could lead to aggregate community-level patterns quite distinct from that of individual species. For example, a mixture of speciesspecic exponential distributions would combine to form a gamma distribution at the community level (Gleser 1989). Changes in movement pattern are likely to alter optimal landscape conguration for pollination services. A wide range of species-specic movement patterns could for example favor a landscape with a broad range of patch sizes as would be the case with a hierarchical or fractal-type landscape. Other areas where the models could be extended to include foraging behavior and population genetics. The inclusion of an optimal foraging module would be interesting as it is known that pollinator foraging activities can be altered by landscape conguration (Steffan-Dewenter and Kuhn 2003) and resource quality (Cartar 2004). Additional factors that might affect outcomes are plant self-fertilization and behavioral resource switching by pollinators, both of which can reduce the strength of Allee effects (Morgan et al. 2005). Positive density dependence can also be enhanced by genetic mechanisms such as self-incompatible alleles (Wagenius et al. 2007), inbreeding depression (Lennartsson 2002), and complementary sex determination as occurs in many Hymenoptera (Zayed and Packer 2005). While the focus of this paper is on spatial patterns and mutualistic feedbacks, it is important to remember that many if not all pollination systems include nectar robbery (Irwin and Brody 1998, Bronstein 2001) and other forms of exploitation (Pellmyr et al. 1996, Yu 2001, Morris et al. 2003) with potentially important consequences for population persistence and pollination services. An important issue not considered here is the phenology of crop and wild plant owering. In the current model, pollinators can survive solely on crop pollen. This is an unrealistic assumption as many crops have relatively brief ower production. Coffee for example blooms over a four-day period (Ricketts 2004) and pollinators must thus forage on other sources for the remainder of the year. With short duration owering, estimates of maximum pollination services reported here

are certainly too large and I would expect to see joint extinction of wild plants and wild pollinators in the vicinity of 50% habitat conversion to croplands (the exact value would depend on details of the specic ecosystem). Future work needs to explore various phenological scenarios to ascertain the effect of owering sequence on potential crop pollination by wild pollinators. Certainly if crop pollen alone is insufcient to maintain wild pollinators, it greatly emphasizes the importance of wild plants in agroecosystems. A nal consideration is the importance of nonequilibrium dynamics. In previously degraded landscapes experiencing natural regrowth or active restoration, facilitating re-invasion and reassembly of the wild pollinator community may be an equal concern to optimizing pollination services. Habitat connectivity may play an important role in restoration efforts. For example, Keitt et al. (2001) showed that a strong Allee effect coupled with habitat fragmentation can halt an invasion when there is insufcient immigration into peripheral patches to overcome positive density dependence (although this effect may be overcome to some extent by the availability of crop pollen). There are likely trade offs between landscape congurations for optimal pollination services and those that promote re-invasion (e.g., the spanning-tree landscape is highly connected, but as a result sacrices some spillover into crops). In cases where agroecosystems may be subjected to aerial spraying or other large scale disturbances, managers should consider, in addition to those that promote crop pollination, landscape designs that will assist in rebuilding pollinator communities to restore wild pollination services as quickly as possible.
ACKNOWLEDGMENTS I thank members of the Restoring an ecosystem service to degraded landscapes: native bees and crop pollination Working Group supported by the National Center for Ecological Analysis and Synthesis, a Center funded by NSF (Grant #DEB-0553768), the University of CaliforniaSanta Barbara, and the State of California. In particular, I am grateful to C. Kremen and N. Williams for their leadership of the group and E. Crone for helpful suggestions during model development and manuscript preparation. I also acknowledge additional support from the David and Lucile Packard Foundation and the Texas Space Grant Consortium funded by NASA. LITERATURE CITED Aguilar, R., L. Ashworth, L. Galetto, and M. A. Aizen. 2006. Plant reproductive susceptibility to habitat fragmentation: review and synthesis through a meta-analysis. Ecology Letters 9:968980. Aizen, M. A., and P. Feinsinger. 1994. Forest fragmentation, pollination, and plant reproduction in a Chaco dry forest, Argentina. Ecology 75:330351. Allen-Wardell, G., et al. 1998. The potential consequences of pollinator declines on the conservation of biodiversity and stability of food crop yields. Conservation Biology 12:817. Amarasekare, P. 2004. Spatial dynamics of mutualistic interactions. Journal of Animal Ecology 73:128142. Ashman, T. L., T. M. Knight, J. A. Steets, P. Amarasekare, M. Burd, D. R. Campbell, M. R. Dudash, M. O. Johnston, S. J.

1572

TIMOTHY H. KEITT

Ecological Applications Vol. 19, No. 6

Mazer, R. J. Mitchell, M. T. Morgan, and W. G. Wilson. 2004. Pollen limitation of plant reproduction: ecological and evolutionary causes and consequences. Ecology 85:2408 2421. Bascompte, J., P. Jordano, C. J. Melian, and J. M. Olesen. 2003. The nested assembly of plantanimal mutualistic networks. Proceedings of the National Academy of Sciences (USA) 100:93839387. Biesmeijer, J. C., S. P. M. Roberts, M. Reemer, R. Ohlemuller, M. Edwards, T. Peeters, A. P. Schaffers, S. G. Potts, R. Kleukers, C. D. Thomas, J. Settele, and W. E. Kunin. 2006. Parallel declines in pollinators and insect-pollinated plants in Britain and the Netherlands. Science 313:351354. Brent, R. P. 1973. Algorithms for minimization without derivatives. Prentice-Hall, Englewood Cliffs, New Jersey, USA. Bronstein, J. L. 2001. The exploitation of mutualisms. Ecology Letters 4:277287. Brosi, B., P. Armsworth, and G. C. Daily. 2008. Optimal design of agricultural landscapes for pollination services. Conservation Letters 1:2736. Cane, J. H., R. L. Minckley, L. J. Kervin, T. H. Roulston, and N. M. Williams. 2006. Complex responses within a desert bee guild (Hymenoptera: Apiformes) to urban habitat fragmentation. Ecological Applications 16:632644. Cartar, R. V. 2004. Resource tracking by bumble bees: responses to plant-level differences in quality. Ecology 85: 27642771. Carvell, C., W. R. Meek, R. F. Pywell, D. Goulson, and M. Nowakowski. 2007. Comparing the efcacy of agri-environment schemes to enhance bumble bee abundance and diversity on arable eld margins. Journal of Applied Ecology 44:2940. Costanza, R., et al. 1997. The value of the worlds ecosystem services and natural capital. Nature 387:253260. Cox-Foster, D. L., et al. 2007. A metagenomic survey of microbes in honey bee colony collapse disorder. Science 318: 283287. Cresswell, J. E., J. L. Osborne, and D. Goulson. 2000. An economic model of the limits to foraging range in central place foragers with numerical solutions for bumblebees. Ecological Entomology 25:249255. Cunningham, S. A. 2000. Depressed pollination in habitat fragments causes low fruit set. Proceedings of the Royal Society B 267:11491152. Donaldson, J., I. Nanni, C. Zachariades, J. Kemper, and J. D. Thompson. 2002. Effects of habitat fragmentation on pollinator diversity and plant reproductive success in renosterveld shrublands of South Africa. Conservation Biology 16:12671276. Fortuna, M. A., and J. Bascompte. 2006. Habitat loss and the structure of plantanimal mutualistic networks. Ecology Letters 9:278283. Gleser, L. J. 1989. The gamma distribution as a mixture of exponential distributions. American Statistician 43:115117. Hastings, A. 2004. Transients: the key to long-term ecological understanding? Trends in Ecology and Evolution 19:3945. Irwin, R. E., and A. K. Brody. 1998. Nectar robbing in Ipomopsis aggregata: effects on pollinator behavior and plant tness. Oecologia 116:519527. Johnson, S. D., and K. E. Steiner. 2000. Generalization versus specialization in plant pollination systems. Trends in Ecology and Evolution 15:140143. Kearns, C. A., D. W. Inouye, and N. M. Waser. 1998. Endangered mutualisms: the conservation of plantpollinator interactions. Annual Review of Ecology and Systematics 29:83112. Keitt, T. H. 2000. Spectral representation of neutral landscapes. Landscape Ecology 15:479494. Keitt, T. H. 2003. Network theory: an evolving approach to landscape conservation. Pages 125134 in V. H. Dale, editor.

Ecological modeling for resource managers. Springer-Verlag, Berlin, Germany. Keitt, T. H., M. A. Lewis, and R. D. Holt. 2001. Allee effects, invasion pinning, and species borders. American Naturalist 157:203216. Kevan, P. G., and T. P. Phillips. 2001. The economic impacts of pollinator declines: an approach to assessing the consequences. Conservation Ecology 5(1):8. Klein, A. M., S. A. Cunningham, M. Bos, and I. SteffanDewenter. 2008. Advances in pollination ecology from tropical plantation crops. Ecology 89:935943. Klein, A. M., I. Steffan-Dewenter, and T. Tscharntke. 2003. Pollination of Coffea canephora in relation to local and regional agroforestry management. Journal of Applied Ecology 40:837845. Klein, A. M., B. E. Vaissiere, J. H. Cane, I. Steffan-Dewenter, S. A. Cunningham, C. Kremen, and T. Tscharntke. 2007. Importance of pollinators in changing landscapes for world crops. Proceedings of the Royal Society B 274:303313. Kohler, F., J. Verhulst, R. van Klink, and D. Kleijn. 2008. At what spatial scale do high-quality habitats enhance the diversity of forbs and pollinators in intensively farmed landscapes? Journal of Applied Ecology 45:753762. Kremen, C., and R. S. Ostfeld. 2005. A call to ecologists: measuring, analyzing, and managing ecosystem services. Frontiers in Ecology and Environment 3:540548. Kremen, C., et al. 2007. Pollination and other ecosystem services produced by mobile organisms: a conceptual framework for the effects of land-use change. Ecology Letters 10:299314. Kremen, C., N. M. Williams, and R. W. Thorp. 2002. Crop pollination from native bees at risk from agricultural intensication. Proceedings of the National Academy of Sciences (USA) 99:1681216816. Lennartsson, T. 2002. Extinction thresholds and disrupted plant-pollinator interactions in fragmented plant populations. Ecology 83:30603072. Losey, J. E., and M. Vaughan. 2006. The economic value of ecological services provided by insects. BioScience 56:311 323. Matson, P. A., W. J. Parton, A. G. Power, and M. J. Swift. 1997. Agricultural intensication and ecosystem properties. Science 277:504509. Matsumura, C., and I. Washitani. 2000. Effects of population size and pollinator limitation on seed-set of Primula sieboldii populations in a fragmented landscape. Ecological Research 15:307322. Morandin, L. A., and M. L. Winston. 2006. Pollinators provide economic incentive to preserve natural land in agroecosystems. AgricultureEcosystemsandEnvironment116:289292. Morgan, M. T., W. G. Wilson, and T. M. Knight. 2005. Plant population dynamics, pollinator foraging, and the selection of self-fertilization. American Naturalist 166:169183. Morris, W. F., J. L. Bronstein, and W. G. Wilson. 2003. Threeway coexistence in obligate mutualistexploiter interactions: the potential role of competition. American Naturalist 161: 860875. Olson, D. M., and F. L. Wackers. 2007. Management of eld margins to maximize multiple ecological services. Journal of Applied Ecology 44:1321. Pauw, A. 2007. Collapse of a pollination web in small conservation areas. Ecology 88:17591769. Pellmyr, O., J. LeebensMack, and C. J. Huth. 1996. Nonmutualistic yucca moths and their evolutionary consequences. Nature 380:155156. Pellmyr, O., and J. N. Thompson. 1996. Sources of variation in pollinator contribution within a build: the effects of plant and pollinator factors. Oecologia 107:595604. Pettis, J., D. Vanengelsdorp, and D. Cox-Foster. 2007. Colony collapse disorder working group pathogen sub-group progress report. American Bee Journal 147:595597.

September 2009

POLLINATION SERVICES IN AGROECOSYSTEMS

1573

Priess, J. A., M. Mimler, A. M. Klein, S. Schwarze, T. Tscharntke, and I. Steffan-Dewenter. 2007. Linking deforestation scenarios to pollination services and economic returns in coffee agroforestry systems. Ecological Applications 17: 407417. Pywell, R. F., E. A. Warman, C. Carvell, T. H. Sparks, L. V. Dicks, D. Bennett, A. Wright, C. N. R. Critchley, and A. Sherwood. 2005. Providing foraging resources for bumblebees in intensively farmed landscapes. Biological Conservation 121:479494. R Development Core Team, 2008. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. hhttp://www. R-project.orgi Rand, T. A., J. M. Tylianakis, and T. Tscharntke. 2006. Spillover edge effects: the dispersal of agriculturally subsidized insect natural enemies into adjacent natural habitats. Ecology Letters 9:603614. Ricketts, T. H. 2004. Tropical forest fragments enhance pollinator activity in nearby coffee crops. Conservation Biology 18:12621271. Ricketts, T., et al. 2008. Landscape effects on crop pollination services: are there general patterns? Ecology Letters 11:499 515. Ringel, M. S., H. H. Hu, G. Anderson, and M. S. Ringel. 1996. The stability and persistence of mutualisms embedded in community interactions. Theoretical Population Biology 50: 281297. Scheffer, M., and S. R. Carpenter. 2003. Catastrophic regime shifts in ecosystems: linking theory to observation. Trends in Ecology and Evolution 18:648656. Seydel, R. 1988. From equilibrium to chaos: practical bifurcation and stability analysis. Elsevier, New York, New York, USA. Steffan-Dewenter, I., and A. Kuhn. 2003. Honeybee foraging in differentially structured landscapes. Proceedings of the Royal Society B 270:569575. Steffan-Dewenter, I., and C. Westphal. 2008. The interplay of pollinator diversity, pollination services and landscape change. Journal of Applied Ecology 45:737741. Stokstad, E. 2007. Entomology: the case of the empty hives. Science 316:970972.

Tilman, D., J. Fargione, B. Wolff, C. DAntonio, A. Dobson, R. Howarth, D. Schindler, W. H. Schlesinger, D. Simberloff, and D. Swackhamer. 2001. Forecasting agriculturally driven global environmental change. Science 292:281284. Tscharntke, T., A. M. Klein, A. Kruess, I. Steffan-Dewenter, and C. Thies. 2005. Landscape perspectives on agricultural intensication and biodiversity: ecosystem service management. Ecology Letters 8:857874. Vazquez, D. P., and M. A. Aizen. 2004. Asymmetric specialization: a pervasive feature of plantpollinator interactions. Ecology 85:12511257. Vitousek, P. M., H. A. Mooney, J. Lubchenco, and J. M. Melillo. 1997. Human domination of Earths ecosystems. Science 277:494499. Wagenius, S., E. Lonsdorf, and C. Neuhauser. 2007. Patch aging and the S-allee effect: breeding system effects on the demographic response of plants to habitat fragmentation. American Naturalist 169:383397. Waser, N. M., L. Chittka, M. V. Price, N. M. Williams, and J. Ollerton. 1996. Generalization in pollination systems, and why it matters. Ecology 77:10431060. Westphal, C., I. Steffan-Dewenter, and T. Tscharntke. 2003. Mass owering crops enhance pollinator densities at a landscape scale. Ecology Letters 6:961965. Wilson, W. G., W. F. Morris, and J. L. Bronstein. 2003. Coexistence of mutualists and exploiters on spatial landscapes. Ecological Monographs 73:397413. Wilson, W. T., and D. M. Menapace. 1979a. Disappearing disease of honey bees: survey of the United States. 1. American Bee Journal 119:118119. Wilson, W. T., and D. M. Menapace. 1979b. Disappearing disease of honey bees: survey of the United States. Conclusion. American Bee Journal 119:184186. Winfree, R., N. M. Williams, J. Dushoff, and C. Kremen. 2007. Native bees provide insurance against ongoing honey bee losses. Ecology Letters 10:11051113. Yu, D. W. 2001. Parasites of mutualisms. Biological Journal of the Linnean Society 72:529546. Zayed, A., and L. Packer. 2005. Complementary sex determination substantially increases extinction proneness of haplodiploid populations. Proceedings of the National Academy of Sciences (USA) 102:1074210746.

APPENDIX A Figures comparing simulation and mathematical model outputs (Ecological Archives A019-062-A1).

APPENDIX B Model sensitivity analysis (Ecological Archives A019-062-A2).

APPENDIX C Pollination services in fragmented landscapes: additional landscapes (Ecological Archives A019-062-A3).